Volume 16, Issue 4 (3-2023)
مرتع 2023, 16(4): 745-764 |
Back to browse issues page
Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
Mahzooni-Kachapi S S, Tahmasebi P, Ebrahimi A, Jouri M H. Evaluation of the ability of different algorithms and visual interpretation of Google Earth images in the separation and classification of plant ecological units. مرتع 2023; 16 (4) :745-764
URL: http://rangelandsrm.ir/article-1-1133-en.html
URL: http://rangelandsrm.ir/article-1-1133-en.html
Department of Range and Watershed Management, Faculty of Natural Resource and Earth Science, Shahrekord University, Shahrekord
Abstract: (1223 Views)
Background and objectives: Satellite images and remote sensing technology are recognized as efficient and modern tools for extracting information related to earth sciences, which make it possible to evaluate and monitor ecosystems at a lower cost than field methods. One of the most important methods of extracting information from satellite data is various image classification techniques. The present study was conducted in order to evaluate the capability of Classification Tree Analysis and Decision Forest algorithms on Sentinel 2 satellite images as well as a visual interpretation of Google Earth images to separate and classify plant ecological units in one of the semi-steppe rangelands of Chaharmahal and Bakhtiari province.
Methodology: In order to distinguish plant ecological units (homogeneous units of vegetation), the visual interpretation method was first used, which included determining polygons on the image and then defining homogeneous areas with similar characteristics in order to identify the dominant type and surface cover of the land. Then, the characteristics of each polygon were interpreted based on the dominant species and the type of disturbances and a general classification was done. After separating the plant ecological units and separating the approximate boundaries of the units, vegetation sampling was done according to the peak growth time of the plant species. After determining the vegetation cover and its production, the average percentage of the estimated vegetation cover in each ecological unit was calculated. For this purpose, first, the dominant plant species of each specific unit and then its accompanying species were determined on the condition of having 50% or more coverage of the previous species. Finally, the identified plant ecological units were named based on the dominant species and by physiognomic and floristic methods and expressed in the form of descriptive statistics. In this study, in addition to the visual interpretation method, Classification Tree Analysis and Decision Forest algorithms were also used to generate a vegetation map. For this purpose, multispectral images of the MSI Sentinel 2 sensor were used as the main processing source. In the next stage, ground control samples were taken randomly from each group of plant ecological units as a model for the spectral characteristics of the classes. Then, the samples were separated into training and experimental datasets, in such a way that a part was divided for classification (one-third of the samples) and another part was divided to estimate the accuracy of the results of the mentioned classification algorithms (two-thirds of the samples). In order to extract information as best as possible, auxiliary layers such as digital elevation model, principal component analysis, and plant indices such as NDVI along with spectral data were used in the classification process. Then the algorithms were classified in Idrisi TerrSet software. In this way, maps of plant ecological units related to the studied area were obtained. In order to evaluate the accuracy of the classification results, the resulting maps were checked with the registered ground reality points. Then, the error matrix related to each method was generated by the software, and finally, the extracted statistics were evaluated and compared.
Results: The results of visual interpretation showed that finally 7 types of plant ecological units were identified that were different in terms of structural features, including Astragalus verus, Bromus tomentellus, Scariola orientalis, Astragalus verus-Bromus tomentellus, Astragalus verus-Stipa hohenikeriana, Bromus tomentellus- Stipa hohenikeriana and Stipa hohenikeriana. The results obtained from the Classification Tree Analysis algorithm showed that the ecological unit Astragalus verus-Stipa hohenikeriana with %0.99 and the ecological unit Bromus tomentellus with %0.90 have the highest and lowest producer accuracy, respectively. While the highest user accuracy is related to the ecological unit Astragalus verus with %0.99 and the lowest value belongs to the ecological unit Stipa hohenikeriana with %0.85. On the other hand, the results of the Decision Forest algorithm indicate producer accuracy and user accuracy above %0.95 in all plant ecological units. So that the ecological unit Astragalus verus and Astragalus verus-Stipa hohenikeriana with %100 and the ecological unit Bromus tomentellus with %0.95 have the highest and lowest producer accuracy, respectively. While the highest user accuracy is related to the ecological unit Astragalus verus-Bromus tomentellus with %100 and the lowest value belongs to the ecological unit Bromus tomentellus-Stipa hohenikeriana with %0.97. The results also show that the overall accuracy and kappa coefficient for the Classification Tree Analysis algorithm is %0.94 and 0.92, respectively, and for the Decision Forest algorithm is %0.99 and 0.97.
Conclusion: According to the obtained results, it can be stated that the Decision Forest algorithm has a higher capability in using spectral information for the classification of plant ecological units in comparison with the Classification Tree Analysis algorithm and led to more accurate results. Also, the integration of the auxiliary bands obtained from the original images along with the raw bands can provide the most valuable information for the identification of plant ecological units. The results of the present research also indicate that if multispectral satellite images with appropriate resolution are not available, using Google Earth images due to their ease of access and their free availability is very suitable and affordable for the preparation of thematic maps such as land cover. It is economical and the map produced from it can be used as a ground reality.
Methodology: In order to distinguish plant ecological units (homogeneous units of vegetation), the visual interpretation method was first used, which included determining polygons on the image and then defining homogeneous areas with similar characteristics in order to identify the dominant type and surface cover of the land. Then, the characteristics of each polygon were interpreted based on the dominant species and the type of disturbances and a general classification was done. After separating the plant ecological units and separating the approximate boundaries of the units, vegetation sampling was done according to the peak growth time of the plant species. After determining the vegetation cover and its production, the average percentage of the estimated vegetation cover in each ecological unit was calculated. For this purpose, first, the dominant plant species of each specific unit and then its accompanying species were determined on the condition of having 50% or more coverage of the previous species. Finally, the identified plant ecological units were named based on the dominant species and by physiognomic and floristic methods and expressed in the form of descriptive statistics. In this study, in addition to the visual interpretation method, Classification Tree Analysis and Decision Forest algorithms were also used to generate a vegetation map. For this purpose, multispectral images of the MSI Sentinel 2 sensor were used as the main processing source. In the next stage, ground control samples were taken randomly from each group of plant ecological units as a model for the spectral characteristics of the classes. Then, the samples were separated into training and experimental datasets, in such a way that a part was divided for classification (one-third of the samples) and another part was divided to estimate the accuracy of the results of the mentioned classification algorithms (two-thirds of the samples). In order to extract information as best as possible, auxiliary layers such as digital elevation model, principal component analysis, and plant indices such as NDVI along with spectral data were used in the classification process. Then the algorithms were classified in Idrisi TerrSet software. In this way, maps of plant ecological units related to the studied area were obtained. In order to evaluate the accuracy of the classification results, the resulting maps were checked with the registered ground reality points. Then, the error matrix related to each method was generated by the software, and finally, the extracted statistics were evaluated and compared.
Results: The results of visual interpretation showed that finally 7 types of plant ecological units were identified that were different in terms of structural features, including Astragalus verus, Bromus tomentellus, Scariola orientalis, Astragalus verus-Bromus tomentellus, Astragalus verus-Stipa hohenikeriana, Bromus tomentellus- Stipa hohenikeriana and Stipa hohenikeriana. The results obtained from the Classification Tree Analysis algorithm showed that the ecological unit Astragalus verus-Stipa hohenikeriana with %0.99 and the ecological unit Bromus tomentellus with %0.90 have the highest and lowest producer accuracy, respectively. While the highest user accuracy is related to the ecological unit Astragalus verus with %0.99 and the lowest value belongs to the ecological unit Stipa hohenikeriana with %0.85. On the other hand, the results of the Decision Forest algorithm indicate producer accuracy and user accuracy above %0.95 in all plant ecological units. So that the ecological unit Astragalus verus and Astragalus verus-Stipa hohenikeriana with %100 and the ecological unit Bromus tomentellus with %0.95 have the highest and lowest producer accuracy, respectively. While the highest user accuracy is related to the ecological unit Astragalus verus-Bromus tomentellus with %100 and the lowest value belongs to the ecological unit Bromus tomentellus-Stipa hohenikeriana with %0.97. The results also show that the overall accuracy and kappa coefficient for the Classification Tree Analysis algorithm is %0.94 and 0.92, respectively, and for the Decision Forest algorithm is %0.99 and 0.97.
Conclusion: According to the obtained results, it can be stated that the Decision Forest algorithm has a higher capability in using spectral information for the classification of plant ecological units in comparison with the Classification Tree Analysis algorithm and led to more accurate results. Also, the integration of the auxiliary bands obtained from the original images along with the raw bands can provide the most valuable information for the identification of plant ecological units. The results of the present research also indicate that if multispectral satellite images with appropriate resolution are not available, using Google Earth images due to their ease of access and their free availability is very suitable and affordable for the preparation of thematic maps such as land cover. It is economical and the map produced from it can be used as a ground reality.
Keywords: Sentinel 2, Classification Tree Analysis Algorithm, Decision Forest Algorithm, auxiliary data
Type of Study: Research |
Subject:
Special
Received: 2022/03/23 | Accepted: 2022/12/3 | Published: 2023/03/1
Received: 2022/03/23 | Accepted: 2022/12/3 | Published: 2023/03/1
Send email to the article author
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |
